Re: [PATCH V3 1/2] topology: Allow multiple entities to provide sched_freq_tick() callback

[Ionela Voinescu]

On Wednesday 17 Feb 2021 at 17:10:27 (+0530), Viresh Kumar wrote:
> On 17-02-21, 11:30, Ionela Voinescu wrote:
> > The problem is not topology_scale_freq_invariant() but whether a scale
> > factor is set for some CPUs.
> > 
> > Scenario (test system above):
> >  - "AMUs" are only supported for [1-2],
> >  - cpufreq_supports_freq_invariance() -> false
> > 
> > What should happen:
> >  - topology_scale_freq_invariant() -> false (passed)
> >  - all CPUs should have their freq_scale unmodified (1024) - (failed)
> >    because only 2 out of 6 CPUs have a method of setting a scale factor
> > 
> > What does happen:
> >  - arch_set_freq_tick() -> topology_set_freq_tick() will set a scale
> >    factor for [1-2] based on AMUs. This should not happen. We will end
> >    up with invariant signals for bigs and signals that are not freq
> >    invariant for littles.
> 
> Another case. cpufreq is included as a module and AMU is implemented
> partially.
> 
> - first time cpufreq driver is inserted, we set up everything and
>   freq_scale gets updated on ticks.
> 
> - remove cpufreq driver, we are back in same situation.
> 

Yes, but the littles (lacking AMUs) would have had a scale factor set
through arch_set_freq_scale() which will correspond to the last
frequency change through the cpufreq driver. When removing the driver,
it's unlikely that the frequency of littles will change (no driver).
 - topology_scale_freq_invariant() will still return true.
 - littles would still  have a scale factor set which is likely accurate
 - bigs will continue updating the scale factor through AMUs.

See a very useful comment someone added recently :) :

"""
+	/*
+	 * We don't need to handle CPUFREQ_REMOVE_POLICY event as the AMU
+	 * counters don't have any dependency on cpufreq driver once we have
+	 * initialized AMU support and enabled invariance. The AMU counters will
+	 * keep on working just fine in the absence of the cpufreq driver, and
+	 * for the CPUs for which there are no counters available, the last set
+	 * value of freq_scale will remain valid as that is the frequency those
+	 * CPUs are running at.
+	 */
"""

> We can't control it that way.. Or we add another call layer in middle
> before the tick-handler gets called for AMU, which will check if we
> are fully invariant or not ?
> 

I would avoid additional work done on the tick, especially for a scenario which
is unlikely. If you think this case is worth supporting, it might be best to do
it at CPUFREQ_REMOVE_POLICY event.

Thanks,
Ionela.

> -- 
> viresh